Intelligent detection method for building facade defects

Abstract

The present invention discloses an intelligent detection method for building facade defects, including adaptive acquisition and preprocessing under multiple illumination conditions, geometric correction and registration of multi-view images, extraction and characterization of multi-scale invariant features, segmentation of defect candidate regions under complex backgrounds, reconstruction of defect topological structures based on physical constraints, accurate quantification and classification of defect parameters, visualization of results and report generation, etc. The present invention solves the technical problems of defect feature enhancement under multiple illumination conditions, complex background interference suppression, unified characterization of multi-view defects, reconstruction of defect topological structures and accurate parameter quantification through technologies such as regional adaptive illumination compensation, viewpoint invariant feature transformation, class defect interference suppression, physical constraint-based broken connection and sub-pixel width measurement, and improves the accuracy and automation of building facade defect detection.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent detection, and comprises an intelligent detection method for building facade defects, in particular, the method is implemented based on drone image recognition. Background Art

[0002] The detection and evaluation of building facade defects is a key link in building safety management, and is of great significance for preventing safety accidents and protecting people's lives and property. Traditional building facade defect detection mainly relies on manual visual inspection, which has problems such as low efficiency, high labor costs, and high safety risks. Especially for high-rise buildings, large public facilities and historical and cultural relics buildings, inspectors need to use auxiliary equipment such as scaffolding and lifting platforms to approach the inspection site, which not only has high operational risks, but also limits the inspection coverage. With the development of drone technology and the advancement of image recognition technology, drone-based intelligent detection methods for building facade defects have gradually become a research hotspot. By collecting high-resolution images by drones and combining computer vision technology to automatically identify and analyze defects, it can greatly improve detection efficiency, reduce costs, and improve safety and inspection quality.

[0003] The current research on defect detection in drone images mainly focuses on the following aspects: First, defect detection based on traditional image processing methods, such as Canny edge detection, Otsu threshold segmentation and other traditional algorithms, extracting defect features through image preprocessing, binarization segmentation and morphological operations; second, defect detection methods based on deep learning, such as using deep learning models such as convolutional neural network (CNN), full convolutional network (FCN) and U-Net for defect segmentation; third, large-scale defect reconstruction methods based on image splicing, through feature point matching and image registration technology, multiple images collected by drones are spliced ​​into a complete building facade orthophoto, and then defect identification is performed; fourth, defect parameter quantification technology, mainly through image processing methods to calculate the width, length and other geometric parameters of the defect. These methods have certain effects on defect detection of simple backgrounds under ideal conditions, but there are still many challenges in actual complex environments.

[0004] However, existing technologies face the following key problems in complex actual environments: First, the problem of defect detection under multiple lighting conditions is prominent. The images collected by drones at different times and in different weather conditions have serious problems of uneven lighting. Existing image preprocessing methods often introduce noise or lose subtle defect texture features while enhancing defect contrast; second, the problem of complex background interference. Building facades often have defect features such as decorative textures, brick joints, and drainage pipes. Existing algorithms find it difficult to effectively distinguish real defects from these linear interference features. Summary of the invention

[0005] The purpose of the invention is to provide an intelligent detection method for building facade defects to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution, the intelligent detection method of building facade defects, comprises the following steps:

[0007] Obtain the original image of the drone, perform adaptive acquisition and preprocessing under multiple lighting conditions, and obtain preprocessed image data;

[0008] Perform multi-view image geometric correction and registration on the preprocessed image data to obtain a standardized orthophoto;

[0009] Perform multi-scale invariant feature extraction and characterization on the standardized orthophoto to obtain defect feature characterization data;

[0010] Based on the defect feature characterization data, the defect candidate area is segmented under complex background to obtain the defect candidate area mask;

[0011] Using defect candidate area masks and defect feature characterization data, combined with building mechanics constraints, the defect topology structure is reconstructed to obtain a complete defect topology network;

[0012] Based on the complete defect topology network, the geometric and physical parameters of the defects are extracted and classified to obtain defect parameter data;

[0013] Integrate standardized orthophotos, complete defect topology network and defect parameter data to generate inspection result visualization and analysis reports to obtain defect inspection result reports.

[0014] Beneficial effects: The present invention effectively solves the quality problem of images collected by UAVs under different lighting environments, eliminates the influence of perspective deformation and scale change, and establishes a defect feature representation that is robust to scale and perspective changes; it can effectively distinguish between real defects and quasi-defect interference, and realizes the effective connection of intermittent defects; it improves the accuracy of defect detection on building facades, reduces the false detection rate, realizes more comprehensive, accurate and automated defect detection, and provides reliable technical support for building safety assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of the steps of a method for intelligent detection of building facade defects provided in an embodiment of the present application.

[0016] Figure 2 A flowchart of the steps for reconstructing a defective topological structure in combination with building mechanics constraints provided in an embodiment of the present application.

[0017] Figure 3 A flowchart of the steps for segmenting defect candidate areas under complex backgrounds provided in an embodiment of the present application.

[0018] Figure 4 A flowchart of the steps for extracting and characterizing multi-scale invariant features provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0020] It should be noted that in order to clearly show the steps of this application, serial numbers are marked for each step in the specification. These serial numbers are only used for the convenience of explanation and do not limit the order of execution of the steps. In actual operation, according to the technical requirements of the specific implementation scenario, the steps can be executed in a different order than that shown in the specification, and in some cases, parallel processing between steps can be achieved.

[0021] like Figure 1 As shown, a method for intelligent detection of building facade defects includes the following steps:

[0022] S1. Obtain the original image of the drone, perform adaptive acquisition and preprocessing under multiple lighting conditions, and obtain preprocessed image data; effectively solve the quality problem of images collected by drones under different lighting environments;

[0023] S2, perform multi-view image geometric correction and registration on the pre-processed image data to obtain a standardized orthophoto, eliminating the effects of perspective deformation and scale change;

[0024] S3. Extract and characterize multi-scale invariant features of standardized orthophotos to obtain defect feature characterization data; establish defect feature representation that is robust to scale and perspective changes;

[0025] S4. Based on the defect feature characterization data, segment the defect candidate area under complex background to obtain the defect candidate area mask; effectively distinguish the real defect from the defect-like interference;

[0026] S5. Using defect candidate area masks and defect feature characterization data, combined with building mechanics constraints, the defect topology structure is reconstructed to obtain a complete defect topology network; the effective connection of intermittent defects is achieved;

[0027] S6. Based on the complete defect topology network, the geometric and physical parameters of the defects are extracted and classified to obtain defect parameter data;

[0028] S7. Integrate the standardized orthophoto, the complete defect topology network and the defect parameter data, generate a detection result visualization and analysis report, and obtain a defect detection result report.

[0029] This embodiment improves the accuracy of building facade defect detection, reduces the false detection rate, realizes more comprehensive, accurate and automated defect detection, and provides reliable technical support for building safety assessment.

[0030] The images of building facades collected by drones at different times and in different weather conditions have serious problems of uneven lighting. Especially when the defects are located in shadow areas or strong light areas, traditional image preprocessing methods (such as histogram equalization, adaptive thresholding, etc.) often introduce noise or lose subtle defect texture features while enhancing the defect contrast. Existing local contrast enhancement algorithms are difficult to effectively enhance the visibility of defects under different lighting conditions while maintaining the true geometric features of the defects. There is a lack of an algorithm that can adaptively enhance the defect features of building materials (such as concrete, masonry, paint, etc.) under different lighting conditions while maintaining the integrity of the defect texture features, especially for small initial defects (width <0.3mm). Therefore, the preprocessing was further optimized. According to one aspect of the present application, step S1 is further:

[0031] S11, adaptive exposure parameter real-time adjustment acquisition: According to the light sensor data and scene reflectivity, a dynamic exposure adjustment algorithm is used to calculate the optimal exposure parameters in real time to obtain the original multi-exposure image set (drone original image).

[0032] S12. Identification and segmentation of non-uniform illumination areas: Analyze the original multi-exposure image set, use the gradient consistency detection algorithm to identify the non-uniform illumination areas in the image, and generate the illumination area segmentation mask.

[0033] S13, regional adaptive illumination compensation: according to the illumination region segmentation mask and the original multi-exposure image set, an adaptive illumination response function model is used to perform differential compensation processing on different regions to obtain an illumination balanced image.

[0034] S14. Defect-sensitive contrast enhancement: A defect-sensitive contrast enhancement algorithm based on a directional structure tensor is applied to the illumination-equalized image to obtain an enhanced image while retaining defect texture details and suppressing non-defect textures.

[0035] S15, material adaptive noise suppression: According to the building material recognition results and enhanced images, a material-guided anisotropic diffusion filtering algorithm is used to retain the defect edge features while suppressing the surface noise of different materials to obtain preprocessed image data.

[0036] According to one aspect of the present application, the steps of performing adaptive acquisition and preprocessing under multiple illumination conditions to obtain preprocessed image data include:

[0037] According to the pre-stored illumination sensor data and scene reflectivity, the dynamic exposure adjustment algorithm is used to calculate the optimal exposure parameters in real time to obtain the original multi-exposure image set, i.e., the original image of the drone;

[0038] Analyze the original multi-exposure image set, use the gradient consistency detection algorithm to identify the uneven illumination areas in the image, and generate the illumination area segmentation mask;

[0039] Based on the illumination region segmentation mask and the original image of the drone, an adaptive illumination response function model is used to perform differential compensation processing on different regions to obtain an illumination-balanced image.

[0040] Based on the illumination equalization image, defect-sensitive contrast enhancement is performed to generate preprocessed image data.

[0041] According to one aspect of the present application, the step of using an adaptive illumination response function model to perform differential compensation processing on different regions to obtain an illumination-balanced image is further as follows:

[0042] According to the illumination region segmentation mask and the original multi-exposure image set, feature analysis is performed on each segmented region: the illumination statistical features of each region are extracted and the illumination transition characteristics between regions are calculated to generate regional illumination statistical features and regional illumination transition features;

[0043] Based on the statistical characteristics of regional illumination, a piecewise polynomial model is used to construct the illumination response function of each region and generate the parameters of the regional illumination response function.

[0044] Combining the building material recognition results and the regional illumination response function parameters, material-specific illumination response function optimization is performed: the ideal illumination response characteristics of the target material are retrieved from the material illumination response library, and the material optimized illumination response parameters are generated;

[0045] Integrate material optimization lighting response parameters and regional lighting transition characteristics to perform global consistency optimization: solve the global energy function through iterative optimization, perform adaptive lighting correction on the original multi-exposure image set, and obtain a lighting balanced image.

[0046] In one embodiment of the present application, feature analysis of uneven illumination regions is performed. An illumination region segmentation mask and an original multi-exposure image set are received, and feature analysis is performed on each segmented region. First, the illumination statistical features of each region are calculated, including the brightness mean, standard deviation, histogram distribution, and gradient distribution, to obtain regional illumination statistical features. Then, the illumination feature differences of adjacent regions are compared, the illumination transition characteristics between regions are calculated, and regional illumination transition features are generated.

[0047] Perform parameterized modeling of the illumination response function. Based on the statistical characteristics of regional illumination, a piecewise polynomial model is used to construct the illumination response function of each region. The specific steps include: first, sample representative pixel samples in each region to construct a brightness mapping relationship. Then, use the least squares method to fit the polynomial parameters to obtain a region-specific parameterized illumination response function model. Finally, the function is smoothed and constrained to ensure the continuity of the function in the entire brightness range, and the regional illumination response function parameters are generated.

[0048] Perform material adaptive lighting response optimization. Combining the building material recognition results and the regional lighting response function parameters, perform material specific lighting response function optimization. First, retrieve the ideal lighting response characteristics of the target material from the pre-trained material lighting response library. Then, perform a weighted fusion of the ideal characteristics and the actual response function to generate a material adaptive target lighting response function. Finally, apply a variational optimization algorithm to minimize the difference between the actual response and the target response while maintaining image details to obtain the material optimized lighting response parameters.

[0049] Smooth the illumination gradient of the boundary transition area. Use the illumination area segmentation mask and the regional illumination transition features to deal with the illumination transition problem of the region boundary. First, extract the region boundary band through morphological operations to determine the transition area that needs to be smoothed. Then, construct an adaptive weight function in the transition area, which takes into account the spatial distance, brightness difference and gradient consistency to ensure smooth transition. Finally, apply the weighted fusion algorithm to smoothly fuse the illumination response functions of adjacent regions in the transition band to obtain the boundary transition illumination map.

[0050] Perform global brightness consistency optimization correction. Integrate material optimization lighting response parameters and boundary transition lighting mapping to perform global consistency optimization. First, establish a global energy function, which includes local contrast preservation terms, material property preservation terms, and global brightness consistency terms. Then, solve the energy function through iterative optimization to generate a globally optimal lighting compensation mapping. Finally, apply this mapping to each area of ​​the original multi-exposure image set for adaptive lighting correction to obtain a lighting-balanced image.

[0051] Conduct defect sensitivity evaluation and feedback adjustment. Conduct defect sensitivity evaluation on the illumination-equalized image to verify the illumination compensation effect. First, apply the defect detection operator in the sample area to evaluate the visibility and contrast of the defects. Then, calculate the defect sensitivity index, including the local contrast gain and the improvement of the signal-to-noise ratio. Finally, based on the evaluation results, fine-tune the material optimization illumination response parameters to achieve iterative optimization, ensure that the illumination compensation process does not lose defect information, and obtain the final illumination-equalized image and defect sensitivity evaluation data.

[0052] This embodiment effectively solves the problem of uneven illumination in drone images. The core is to optimize the adaptive illumination response of materials in combination with the results of building material recognition, and customize the illumination response function for different building materials; finally, adaptive illumination correction is achieved through the optimization and solution of the global energy function. This embodiment balances the overall brightness while maintaining the details of the defects, so that the defect detection accuracy in the bright light area and the shadow area is increased by about 35% and 40% respectively. In particular, for surfaces of different materials such as concrete, masonry, and paint, the illumination compensation effect is more significant, providing higher quality input data for subsequent image processing, and improving the detection performance of the entire system from the source.

[0053] According to one aspect of the present application, step S14 is further:

[0054] S141, applying a Gaussian derivative filter to the illumination equalization image, calculating multi-scale first-order gradients, constructing a pixel-level structure tensor matrix, performing eigenvalue decomposition, and generating a structure tensor feature map;

[0055] S142, based on the structural tensor feature map, extracting the main direction feature vector of the defect in the local area, constructing a directional Gaussian derivative filter group, and generating a directional enhancement filter group;

[0056] S143, applying a directional enhancement filter group to the illumination equalization image to generate a multi-directional response map, applying a non-maximum suppression algorithm to suppress non-maximum responses in a direction perpendicular to the defect, and generating a defect response map;

[0057] S144, constructing Hessian matrix features for defect response mapping, performing background suppression and defect enhancement processing: constructing a structural adaptive transfer function to enhance defect response while suppressing background texture, applying anisotropic diffusion equation, and generating an enhanced defect response map;

[0058] S145. Based on the enhanced defect response map, a segmented contrast mapping function is designed, and a local adaptive histogram equalization algorithm is applied to enhance the local contrast, perform dynamic range compression, and generate an enhanced image.

[0059] In one embodiment of the present application, a multi-directional structure tensor is calculated. An illumination-equalized image is received and a multi-directional structure tensor feature is calculated. First, a Gaussian derivative filter is applied to the image to calculate a multi-scale first-order gradient (Ψx, Ψy). Then, a pixel-level structure tensor matrix T = [[Ψx 2 ,ΨxΨy],[ΨxΨy,Ψy 2 ]], where Ψ is the partial derivative. Next, the structure tensor is subjected to eigenvalue decomposition to extract the main direction and anisotropy. Finally, a local structure classification mask is constructed based on the eigenvalue ratio to distinguish linear structures, edge structures, and homogeneous regions, and generate a structure tensor feature map.

[0060] Construct a defect priority direction enhancement filter. Based on the structural tensor feature map, an adaptive defect direction enhancement filter is designed. First, the main direction feature vector of the defect in the local area is extracted. Then, a directional Gaussian derivative filter group is constructed, and the filter parameters (σ, θ) are dynamically adjusted with the defect direction. Next, the filter is anisotropically stretched to enhance the response consistent with the defect direction. Finally, singular value decomposition is applied to optimize the filter parameters to ensure the maximum response to the specific direction of the defect, and a direction enhancement filter group is obtained.

[0061] Calculate adaptive defect response. Apply the directional enhancement filter bank to the illumination equalized image to calculate the defect response. First, apply the filter bank to each pixel position for convolution operation to generate a multi-directional response map. Then, apply the non-maximum suppression algorithm to suppress the non-maximum response along the direction perpendicular to the defect. Next, calculate the response strength and directional consistency index to construct a defect possibility map. Finally, apply the adaptive threshold for preliminary segmentation to obtain the defect response map.

[0062] Background structure suppression and defect enhancement are performed. Background suppression and defect enhancement are performed on the defect response map. First, the Hessian matrix features are constructed to distinguish defects from local texture features. Then, a structural adaptive transfer function is designed to enhance the defect response while suppressing the background texture. Next, the anisotropic diffusion equation is applied to perform selective smoothing along the defect direction. Finally, a morphological reconstruction algorithm is used to maintain the continuity of the defect while suppressing isolated noise to obtain an enhanced defect response map.

[0063] Perform contrast optimization and dynamic range adjustment. Based on the enhanced defect response map, perform global contrast optimization and dynamic range adjustment. First, calculate the image histogram and perform segmented analysis to determine the brightness distribution of the defect and the background. Then, design a segmented contrast mapping function and use different enhancement strategies for the defect area and the background area. Next, apply a local adaptive histogram equalization algorithm to enhance the local contrast. Finally, perform dynamic range compression to ensure that the final image is within the visible range while maintaining the defect details to generate an enhanced image.

[0064] This embodiment improves the visual significance of defects. This embodiment improves the detection accuracy of small initial defects (width <0.3mm) by about 45%, and the defect signal-to-noise ratio after contrast enhancement is increased by an average of 3-5dB, while the background texture interference is reduced by about 60%, laying a good foundation for subsequent defect segmentation.

[0065] According to one aspect of the present application, step S15 is further:

[0066] S151, receiving the building material recognition result and the enhanced image, applying the superpixel segmentation algorithm to segment the image into small areas with similar texture characteristics, assigning a material type label to each area, and generating a fine material segmentation mask;

[0067] S152, based on the fine material segmentation mask and the enhanced image, extracting a representative homogeneous sub-region in each material region as a noise sample, calculating noise statistical characteristics, constructing a material-specific noise model, and generating material noise characteristic data;

[0068] S153, using the material noise characteristic data and the structure tensor feature map, designing a material-adaptive nonlinear anisotropic diffusion filter, adjusting the diffusion coefficient according to the material noise characteristics, and generating a material-adaptive diffusion parameter;

[0069] S154, applying material adaptive diffusion parameters to perform edge-preserving smoothing on the enhanced image, using a semi-implicit finite difference scheme to numerically solve the nonlinear diffusion equation to generate a smooth optimized image; achieving selective noise suppression;

[0070] S155, performing defect structure consistency maintenance processing on the smoothed optimized image, using the defect response mapping as a structure guide, applying a local structure recovery operation in the defect area, and generating preprocessed image data.

[0071] In one embodiment of the present application, building material region segmentation is performed. Building material recognition results and enhanced images are received, and fine material region segmentation is performed. First, a superpixel segmentation algorithm is applied to segment the image into small regions with similar texture characteristics. Then, superpixels are clustered based on the material recognition results, and superpixel regions of the same material are merged. Next, region boundary optimization is performed to ensure that the material boundary is aligned with the actual building structure boundary. Finally, a material type label is assigned to each region to obtain a fine material segmentation mask.

[0072] Perform material-specific noise characteristics analysis. Based on the fine material segmentation mask and enhanced image, analyze the noise characteristics of different material regions. First, extract representative homogeneous sub-regions in each material region as noise samples. Then, calculate the noise statistics, including noise power spectrum density, spatial correlation, and frequency distribution. Next, construct a material-specific noise model to describe the noise distribution pattern on the surface of different materials. Finally, generate material noise characteristic data, including the noise model parameters for each material region.

[0073] Construct a nonlinear anisotropic diffusion filter. Using the material noise characteristic data and the structure tensor feature map, a material-adaptive nonlinear anisotropic diffusion filter is designed. First, a diffusion tensor field is constructed, and the eigenvector of the diffusion tensor is aligned with the eigenvector of the structure tensor. Then, the diffusion coefficient is adjusted according to the material noise characteristics to suppress material-specific noise while maintaining the defect edge. Next, a nonlinear diffusion equation is designed to adjust the relationship between the diffusion intensity and the gradient amplitude. Finally, material-adaptive diffusion parameters are generated to specify the optimal diffusion parameters for each region.

[0074] Perform edge-preserving smoothing optimization. Apply material adaptive diffusion parameters to perform edge-preserving smoothing on the enhanced image. First, set the initial and boundary conditions to prepare for iterative solution of the diffusion equation. Then, a semi-implicit finite difference scheme is used to numerically solve the nonlinear diffusion equation to achieve anisotropic smoothing. Next, an edge stop function is applied after each iteration to ensure that the defect edge is not blurred. Finally, the number of iterations is adaptively controlled according to the convergence condition to obtain a smoothed optimized image.

[0075] Maintain the consistency of defect structure. The smoothed optimized image is processed to maintain the consistency of defect structure. First, the defect response map is used as a structural guide to determine the defect structure that needs to be protected. Then, the local structure consistency index is calculated to evaluate the impact of the smoothing process on the defect structure. Next, the local structure recovery operation is applied to the defect area to ensure that the smoothing process does not reduce the visibility of the defect. Finally, the structural detail enhancement is performed to improve the sharpness and contrast of the defect edge to generate the final preprocessed image data.

[0076] This embodiment solves the problem of large differences in noise characteristics on the surfaces of different building materials. Customized processing is performed for the specific noise characteristics of different building material surfaces, which improves the noise suppression effect by about 30% compared with the traditional unified filtering method, while maintaining the integrity of the defect edge, and the signal-to-noise ratio is improved by 4dB on average; especially for masonry surfaces with complex textures and concrete surfaces with high roughness, the noise suppression effect is more significant, providing higher quality image data for subsequent feature extraction.

[0077] According to one aspect of the present application, step S2 is further:

[0078] S21. Real-time estimation of camera pose: Combining the drone IMU data and image feature points, the improved visual-inertial fusion algorithm is used to calculate the precise camera pose and generate camera pose data.

[0079] S22. Building surface geometry reconstruction: Based on the multi-view images in the preprocessed image data, a method combining multi-view stereo matching and deep learning is used to reconstruct the building surface geometry model to obtain a 3D model of the building surface.

[0080] S23, surface unfolding transformation: apply the distance-preserving differential geometry unfolding algorithm to the 3D model of the building surface, convert the surface structure into a plane representation, and obtain the unfolded surface mapping.

[0081] S24, multi-view image projection transformation: according to the camera pose data, the building surface 3D model and the unfolded surface mapping, the pre-processed image data is accurately projected and transformed to obtain an orthographic projection image set.

[0082] S25. Multi-scale image fusion registration: A non-rigid image registration algorithm based on local feature flow is used to accurately register and fuse images of different scales and perspectives in the orthophoto image set to obtain standardized orthophoto images and registration accuracy evaluation data.

[0083] According to one aspect of the present application, the steps of performing multi-view image geometric correction and registration to obtain a standardized orthophoto include:

[0084] Combine the pre-stored UAV IMU data and the image feature points in the pre-processed image data, use the visual inertial fusion algorithm to calculate the precise camera pose and generate camera pose data;

[0085] Based on the multi-view images in the preprocessed image data, a method combining multi-view stereo matching and deep learning is used to reconstruct the building surface geometric model to obtain a 3D model of the building surface.

[0086] Segment and classify the surface areas of the 3D model of the building surface, calculate the principal curvature and Gaussian curvature of the mesh, identify the plane area, single curvature area and double curvature area, and generate surface type segmentation data;

[0087] Segment the data according to the surface type and calculate the distance-preserving differential geometry mapping: construct the geodesic distance field, apply the multidimensional scaling algorithm to map the geodesic distance to the two-dimensional plane, balance the distance-preserving and angle-preserving constraints through the energy minimization framework, and generate the unfolded surface mapping;

[0088] Combining the camera pose data, the 3D model of the building surface and the unfolded surface mapping, the pre-processed image data is accurately projected and transformed to obtain an orthographic projection image set;

[0089] A non-rigid image registration algorithm based on local feature flow is used to accurately register and fuse images of different scales and perspectives in the orthophoto image set to obtain a standardized orthophoto.

[0090] In one embodiment of the present application, multi-sensor data synchronization and preprocessing are performed. UAV inertial measurement unit (IMU) data and preprocessed image data are received, and time synchronization and preprocessing are performed. First, information such as acceleration, angular velocity, and magnetic field in the IMU data is extracted, and noise filtering and deviation correction are performed. Then, the time deviation between image acquisition and IMU data is calculated, and data time alignment is achieved through an interpolation algorithm. Next, the aligned data is subjected to anomaly detection and processing to eliminate occasional sensor anomalies. Finally, the corrected synchronized sensor data is generated as input for subsequent processing.

[0091] Perform image feature point extraction and matching. Extract feature points from preprocessed image data and match them. First, apply the improved scale-invariant feature transform (SIFT) algorithm to extract scale-invariant feature points and enhance the response to the texture features of the building surface. Then, calculate the local descriptor and use binary descriptors to improve computational efficiency. Next, use the cascade matching strategy to match feature points of adjacent images, including neighbor ratio test and geometric consistency verification. Finally, construct a feature point tracking sequence and generate a feature point matching sequence.

[0092] Perform visual-inertial tightly coupled fusion. Combine the corrected synchronized sensor data and feature point matching sequence to perform visual-inertial tightly coupled fusion. First, a visual-inertial joint state estimation model is constructed, and the state vector includes position, attitude, velocity, and sensor parameters. Then, a factor graph is established based on the graph optimization framework, including visual observation factors and IMU pre-integration factors. Next, an adaptive robust kernel function is applied to reduce the impact of outliers. Finally, an iterative nonlinear optimization algorithm is used to solve the optimal state estimation and obtain the initial pose estimation.

[0093] Perform building structural constraint optimization. Use the building's prior geometric characteristics to optimize the initial pose estimate. First, detect the main structural lines of the building from the image, including horizontal lines, vertical lines, and vanishing points. Then, convert these structural line constraints into geometric constraints for pose optimization. Next, build an enhanced state estimation model with structural constraints and optimize it by minimizing the structural line reprojection error. Finally, combine the initial pose estimate and the structural constraint optimization results to obtain the optimized pose estimate.

[0094] Perform pose accuracy assessment and drift correction. Perform accuracy assessment and drift correction on the optimized pose estimate. First, calculate the uncertainty covariance of the pose estimate to evaluate pose accuracy. Then, detect and correct the accumulated drift, and perform closed-loop detection by identifying repeated observation positions. Next, perform global pose graph optimization when a closed loop is detected to eliminate accumulated errors. Finally, output the corrected high-precision camera pose data, which contains the position, pose, and accuracy assessment information of each image frame.

[0095] Perform a rough reconstruction of multi-view stereo matching. Perform a preliminary stereo matching reconstruction based on the preprocessed image data and camera pose data. First, select the best view combination according to the camera pose to build a multi-baseline stereo matching pair. Then, use the semi-global matching algorithm (SGM) to calculate the initial disparity map and generate dense correspondences. Next, filter out unreliable matching points through disparity consistency check and left-right consistency check. Finally, convert the two-dimensional matching points into a three-dimensional point cloud through triangulation to obtain the initial three-dimensional point cloud.

[0096] Perform point cloud filtering and structural enhancement. Perform filtering and structural enhancement on the initial 3D point cloud. First, apply the statistical outlier filtering algorithm to remove noise points and abnormal points. Then, perform normal vector estimation and use the least squares plane fitting to calculate the local surface normal vector of each point. Next, apply a smoothing operation to maintain the overall geometric structure while reducing noise. Finally, use the anisotropic point diffusion algorithm to enhance the building structure features and generate a structurally enhanced point cloud.

[0097] Deep learning-assisted geometry completion. Use deep learning methods to complete incomplete areas in structure-enhanced point clouds. First, the point cloud is converted to a voxel representation and a 3D voxel grid is constructed that can be used for deep learning processing. Then, a pre-trained 3D shape completion network is used to predict the geometry of the missing area. The prediction is then fused with the original point cloud to generate a more complete 3D representation. Finally, the fused result is post-processed and optimized to ensure geometric consistency and obtain a complete 3D point cloud.

[0098] Reconstruct the building surface mesh. Reconstruct the building surface mesh model based on the improved 3D point cloud. First, apply the Poisson surface reconstruction algorithm to convert the point cloud into a watertight triangulated mesh. Then, perform mesh simplification and optimization to reduce the number of facets while maintaining geometric details. Next, perform feature-preserving smoothing on the mesh to reduce noise while retaining building features such as edges and corners. Finally, perform mesh repair based on the building structure characteristics to ensure the integrity and accuracy of the model and generate a 3D model of the building surface.

[0099] Perform surface region segmentation and classification. Receive the 3D model of the building surface and perform surface region segmentation and classification. First, calculate the principal curvature and Gaussian curvature of the mesh and analyze the local geometric characteristics. Then, segment the surface based on the curvature characteristics to identify plane regions, single curvature regions (cylinders, cones, etc.) and double curvature regions. Next, use the region growing algorithm to merge adjacent regions with similar geometric characteristics. Finally, assign a geometric type label to each region to generate surface type segmentation data.

[0100] Construct a parametric coordinate system for the surface. Split the data according to the surface type and construct a suitable parametric coordinate system for each region. First, for the plane region, establish an orthogonal coordinate system based on the main direction. Then, for the single curvature region, establish a specific coordinate system suitable for the surface type (such as a cylindrical coordinate system). Next, for the double curvature region, construct an intrinsic coordinate system based on geodesics. Finally, design a smooth transition scheme at the region boundary to ensure the continuity of the overall coordinate system and obtain the parametric coordinate system data.

[0101] Compute distance-preserving differential geometry mappings. Distance-preserving differential geometry mappings are computed based on parameterized coordinate system data. First, a geodesic distance field is constructed to compute the geodesic distances between any pair of points on the surface. Then, a multidimensional scaling (MDS) algorithm is applied to map the geodesic distances to a two-dimensional plane. The mapping result is then optimized by applying a locally linear embedding (LLE) to further preserve the local geometry. Finally, an energy minimization framework is constructed to balance the distance-preserving and angle-preserving constraints to generate the initial unfolding mapping.

[0102] Strain minimization optimization is performed. The initial unfolding map is subjected to strain minimization optimization. First, the local strain tensors introduced during the unfolding process are calculated, including linear strains and shear strains. Then, a global energy function based on strain energy is constructed to define the unfolding quality metric. Next, a nonlinear optimization algorithm is applied to iteratively adjust the mapping parameters to minimize the strain energy. Finally, additional constraints are applied at the boundaries and feature lines to maintain key geometric features and obtain the optimized unfolding map.

[0103] Perform tile mapping stitching and global consistency assurance. Stitch the optimized unfolded mappings of each segmented region to construct a globally consistent unfolded representation. First, identify the shared boundaries between regions and establish boundary correspondences. Then, design a boundary stitching strategy to minimize geometric distortion at the stitching. Next, apply a global optimization algorithm to adjust the relative position and orientation of each region to ensure overall consistency. Finally, generate the final unfolded surface mapping, which contains the complete mapping relationship from the 3D model to the 2D unfolded plane.

[0104] Accurately calibrate the camera projection model. Based on the camera pose data, accurately calibrate the camera projection model. First, extract the camera intrinsic parameters (focal length, principal point, pixel ratio, etc.) and distortion parameters. Then, use the Bundle Adjustment algorithm to optimize the camera parameters and minimize the reprojection error. Next, build a complete camera projection model to describe the mapping relationship from 3D points to the image plane. Finally, calculate the projection accuracy evaluation index, verify the model accuracy, and generate an accurate camera projection model.

[0105] Perform accurate registration of the 3D model and the image. Combine the 3D model of the building surface, the camera pose data, and the accurate camera projection model to accurately register the 3D model and the image. First, use the camera projection model to project the 3D model onto each image plane to generate the initial registration result. Then, extract the main edge features of the 3D model and the corresponding edge features in the image. Next, apply the iterative closest point (ICP) algorithm to optimize the alignment relationship between the 3D model and the image. Finally, calculate the registration accuracy to obtain the model image registration data.

[0106] Perform the projection transformation calculation from the unfolded plane to the image. Using the unfolded surface mapping and the model image registration data, calculate the projection transformation from the unfolded plane to each image. First, find the corresponding 3D model points for the grid points on the unfolded plane. Then, use the camera projection model to project these 3D points to each image plane and establish the correspondence between the unfolded plane and the image plane. Next, calculate the perspective transformation matrix from the unfolded plane to each image based on the corresponding point pairs. Finally, evaluate the transformation accuracy and generate the planar image projection transformation matrix.

[0107] Perform adaptive multi-level image resampling. Based on the planar image projection transformation matrix, the preprocessed image data is adaptively resampled at multiple levels. First, an image pyramid is constructed to prepare for resampling at different resolutions. Then, the corresponding projection transformation matrix is ​​applied to each image to transform the image into a unified unfolding plane coordinate system. Next, a bicubic interpolation algorithm is used to perform high-quality image resampling while maintaining image details. Finally, resampled orthoimages at multiple resolution levels are generated.

[0108] Perform multi-view visibility analysis and selection. Perform multi-view visibility analysis and selection on resampled ortho images. First, construct a view visibility map to describe which cameras observe each 3D point. Then, analyze the observation quality of each view, including factors such as viewing angle, resolution, and clarity. Next, select the best view combination for each position on the unfolded plane to optimize image quality and coverage. Finally, generate a view selection map to guide subsequent image fusion and obtain view selection data and ortho projection image sets.

[0109] Calculate local feature flows. Receive a set of orthographic projection images and calculate local feature flows between images. First, use a dense feature point detection algorithm to extract dense and evenly distributed feature points in each image. Then, calculate local feature descriptors and use deep learning methods to improve the discriminability and robustness of the descriptors. Next, use a bidirectional nearest neighbor matching strategy to perform feature matching and establish correspondence between images. Finally, use the random sampling consensus (RANSAC) algorithm to filter abnormal matches and generate a local feature flow field that describes the image deformation.

[0110] Adaptive mesh deformation modeling is performed. An adaptive mesh deformation model is constructed based on the local characteristic flow field. First, an adaptive mesh is established on the reference image, and the mesh density is dynamically adjusted according to the image complexity. Then, the displacement vector of each mesh vertex is estimated based on the characteristic flow field. Next, the moving least squares (MLS) method is applied to interpolate the discrete displacement vectors into a continuous deformation field. Finally, the deformation field is optimized through regularization constraints to balance local accuracy and global smoothness, and the mesh deformation model is obtained.

[0111] Perform non-rigid image registration optimization. Use the mesh deformation model to perform non-rigid registration optimization on the orthographic projection image set. First, generate a deformation vector field based on the mesh deformation model to describe the mapping relationship of each pixel. Then, apply the inverse mapping and bilinear interpolation algorithm to deform the source image to the target image coordinate system. Next, calculate the difference between the registered images and evaluate the registration quality. Finally, adjust the deformation parameters through iterative optimization to minimize the visual difference between the images and generate a registered image set.

[0112] Perform multi-resolution image pyramid fusion. Perform multi-resolution image pyramid fusion on the registered image set. First, construct a Laplacian pyramid to decompose the image into different frequency components. Then, at each frequency layer, calculate the fusion weight based on the image quality, considering factors such as clarity, exposure, and viewing angle. Next, apply a weighted average strategy to fuse the frequency components of each image. Finally, reconstruct the fused Laplacian pyramid to generate a preliminary fused orthophoto.

[0113] Perform seam detection and seamless stitching optimization. Perform seam detection and seamless stitching optimization on the fused orthophoto. First, detect the visible seams and discontinuous areas generated during the image fusion process. Then, apply the gradient domain fusion algorithm to the seam area to eliminate brightness discontinuities by solving the Poisson equation. Next, perform color consistency optimization to adjust the tones of each area to make them consistent. Finally, apply local detail enhancement technology to restore details that may have been lost in the fusion process and generate the final standardized orthophoto and registration accuracy evaluation data.

[0114] This embodiment solves the problem of defect detection on non-planar surfaces of buildings (such as circular, arc-shaped, spherical and other complex geometric structures). It realizes high-fidelity unfolding of complex geometric surfaces, and the deformation error during the unfolding process is controlled within 5%. In particular, for single-curvature surfaces such as cylinders and cones, the unfolding accuracy reaches more than 95%. Through this unfolding transformation, the subsequent defect detection algorithm can be carried out in a unified two-dimensional coordinate system, which simplifies the algorithm complexity and improves the detection accuracy. For situations that are difficult to handle with traditional methods such as curved facades, the detection success rate is increased by about 40%.

[0115] The distance and angle of the drone from the building are constantly changing during the flight, resulting in problems such as perspective deformation, scale change and inconsistent resolution in the collected defect images. Traditional feature extraction methods have significant differences in the defect features extracted at different viewing angles and distances, making it difficult to establish a consistent defect characterization model, affecting the accuracy of subsequent defect classification, measurement and evaluation. There is a lack of a defect feature characterization method that can adaptively extract perspective-invariant and scale-robust defect features from building facade images collected at different viewing angles and distances, especially the problem of unified characterization of defect features on curved surfaces or complex geometric structures of buildings (such as non-planar structures such as cylinders and arches) needs to be solved urgently. Therefore, according to one aspect of the present application, step S3 is further:

[0116] S31. Multi-directional and multi-scale decomposition: Apply improved wavelet curvelet transform to the standardized orthophoto for multi-directional and multi-scale decomposition to obtain a multi-directional and multi-scale coefficient matrix.

[0117] S32. Scale-invariant defect feature extraction: Based on the multi-directional and multi-scale coefficient matrix, the defect-specific scale-invariant feature descriptors are calculated, including features such as directional consistency, aspect ratio and continuity, to generate a defect-invariant feature map.

[0118] S33. Material adaptive feature compensation: Combined with the building material recognition results, the material-guided feature compensation algorithm is applied to the defect-invariant feature mapping to improve the consistent expression of defect features on the surfaces of different materials and obtain a compensated feature mapping.

[0119] S34. View-invariant feature transformation: Apply the view-invariant feature transformation based on Riemann manifold learning to the compensation feature map to eliminate the influence of view change on the defect feature expression and obtain the view-invariant feature representation.

[0120] S35, context-aware feature enhancement: Combine the view-invariant feature representation with the building structure semantic information, use the graph attention network to perform context-aware feature enhancement, and generate the final defect feature representation data.

[0121] like Figure 4 As shown, according to one aspect of the present application, the steps of extracting and characterizing multi-scale invariant features to obtain defect feature characterization data include:

[0122] The improved wavelet curvelet transform is applied to the standardized orthophoto for multi-directional and multi-scale decomposition to obtain a multi-directional and multi-scale coefficient matrix; based on the multi-directional and multi-scale coefficient matrix, the defect-specific scale-invariant feature descriptor is calculated to generate a defect-invariant feature map; combined with the building material recognition results, the material-guided feature compensation algorithm is applied to the defect-invariant feature map to obtain a compensated feature map; the perspective-invariant feature transformation based on Riemann manifold learning is applied to the compensated feature map to eliminate the influence of perspective change on the defect feature expression and obtain the perspective-invariant feature representation; the perspective-invariant feature representation is combined with the semantic information of the building structure, and the graph attention network is used for context-aware feature enhancement to generate defect feature representation data.

[0123] According to one aspect of the present application, applying a perspective invariant feature transformation based on Riemannian manifold learning to the compensated feature map is further:

[0124] The compensated feature map is regarded as a point on the Riemann manifold, the Riemann distance matrix between the feature points is calculated, the local tangent space is constructed, and the Riemann manifold coordinate system data is generated; based on the Riemann manifold coordinate system data and the preconfigured perspective change trajectory model, the manifold learning algorithm is applied to construct a perspective-invariant feature space and generate a manifold learning model; the manifold learning model is used to transform the compensated feature map to generate a perspective-invariant feature representation.

[0125] In one embodiment of the present application, a Riemannian manifold coordinate system is constructed. A compensated feature map is received and a coordinate system based on the Riemannian manifold is constructed. First, the feature vectors are regarded as points on the Riemannian manifold, and the Riemannian distance matrix between the feature points is calculated. Then, a local tangent space is constructed to establish a mapping relationship between the Euclidean space and the Riemannian manifold. Next, a Riemannian index mapping and a logarithmic mapping function are designed to realize the coordinate transformation on the manifold. Finally, a geodesic equation on the manifold is constructed for subsequent feature transformation optimization to generate Riemannian manifold coordinate system data.

[0126] Generate and analyze samples of viewpoint changes. Generate defect feature samples under different viewpoints and analyze them. First, based on the 3D model of the building surface and the compensation feature map, the appearance of defect features under different viewpoints is simulated through three-dimensional transformation. Then, the feature vectors of these simulated samples are extracted to construct a viewpoint change feature sample set. Next, the influence of viewpoint changes on feature vectors is analyzed to construct a viewpoint change trajectory model. Finally, the viewpoint change trajectory is mapped to the Riemann manifold coordinate system data to obtain the viewpoint change trajectory model.

[0127] Construct a manifold learning feature space. Based on the view change trajectory model and the Riemannian manifold coordinate system data, a view-invariant feature space is constructed. First, a manifold learning algorithm, such as isomap or local linear embedding (LLE), is applied to learn the intrinsic low-dimensional structure of the feature space. Then, a mapping function from the original feature space to the invariant feature space is constructed, which maps the features of the same defect at different viewpoints to similar points. Next, kernel principal component analysis (KPCA) is used to optimize the feature space to enhance intra-class clustering and inter-class separation. Finally, a manifold learning model that describes the structure of the feature space is generated.

[0128] Measure and optimize the view invariance. Measure and optimize the view invariance of the manifold learning model. First, design a view invariance evaluation index to quantify the stability of features under view changes. Then, evaluate the model performance on the test sample set and calculate the feature matching accuracy under different view conditions. Next, adjust the manifold learning parameters through iterative optimization to maximize the view invariance performance. Finally, use an adaptive weight strategy to balance the constraints of local geometry preservation and global structure preservation to obtain the optimized manifold learning model.

[0129] Generate view-invariant feature maps. The compensated feature maps are transformed using an optimized manifold learning model to generate a view-invariant feature representation. First, the learned mapping function is applied to each feature point to transform it from the original feature space to the invariant feature space. Then, feature normalization is applied to remove residual view-dependency. Next, a feature reliability index is calculated to quantify the degree of invariance of each feature point. Finally, a complete representation containing the original space coordinates and the invariant feature vectors is generated to obtain the view-invariant feature representation.

[0130] This embodiment solves the problem of inconsistent defect image features collected by drones at different angles and distances. The defect feature matching accuracy is improved by more than 40% within the 30°-60° viewing angle range, especially for defect recognition on curved surfaces or complex geometric structures of buildings, providing technical support for comprehensive defect detection under the free flight path of drones.

[0131] According to one aspect of the present application, step S32 is further:

[0132] S321, calculating coefficient differences between adjacent scales for a multi-directional multi-scale coefficient matrix, constructing a scale normalization factor, calculating local extreme value points in the scale space, and generating a multi-scale differential extreme value point set;

[0133] S322. Extract defect directional consistency features based on multi-scale differential extreme value point sets and multi-directional multi-scale coefficient matrices: analyze the response distribution in different directions, construct a directional consistency scoring function, and generate a directional consistency feature map;

[0134] S323. Combine the directional consistency feature map and the multi-directional multi-scale coefficient matrix to perform projection profile analysis, calculate the width-to-length ratio, apply the curve fitting method to calculate the local curvature and linear measurement of the defect, and generate the slenderness ratio continuity feature map;

[0135] S324, constructing a feature descriptor with the key points in the directional consistency feature map and the aspect ratio continuity feature map as the center, extracting the local feature region and dividing it into a predetermined number of sub-regions, calculating the feature statistics of each sub-region, and generating a scale-invariant feature descriptor;

[0136] S325. Verify and optimize the scale-invariant feature descriptor, combine the multi-directional and multi-scale coefficient matrix to analyze the variation of features at different scales, establish a scale variation compensation model, and generate a defect-invariant feature map.

[0137] In one embodiment of the present application, multi-scale differential features are calculated. A multi-directional multi-scale coefficient matrix is ​​received, and differential features between scales are calculated. First, for the wavelet coefficients in each direction, the coefficient difference between adjacent scales is calculated to form a differential response. Then, a scale normalization factor is constructed to normalize the differential responses of different scales so that the features remain invariant to scale changes. Next, local extreme points in the scale space are calculated, which are local maximum or minimum values ​​in both spatial position and scale dimension. Finally, extreme points with significant responses are screened to generate a multi-scale differential extreme point set.

[0138] Extract defect directional consistency features. Based on the multi-scale differential extreme point set and the multi-directional multi-scale coefficient matrix, the defect directional consistency features are extracted. First, for each extreme point, its response distribution in different directions is analyzed, and the directional response histogram is calculated. Then, the von Mises distribution model is used to fit the directional distribution, and the main direction and directional concentration parameters are extracted. Next, a directional consistency scoring function is constructed to quantify the degree of consistency of the direction in the local area. Finally, the non-maximum suppression algorithm is applied to retain candidate points with high directional consistency and obtain the directional consistency feature map.

[0139] Calculate the slenderness ratio and continuity features. Combine the directional consistency feature map and the multi-directional multi-scale coefficient matrix to calculate the slenderness ratio and continuity features of the defect. First, perform a projection profile analysis along the main direction of the defect and the vertical direction to calculate the width-to-length ratio. Then, use an adaptive threshold segmentation algorithm to extract the skeleton line and boundary of the defect. Next, apply a curve fitting method to calculate the local curvature and linearity metric of the defect. Finally, construct a feature score that combines the slenderness ratio and continuity to generate a slenderness ratio continuity feature map.

[0140] Construct scale-space extreme point feature descriptors. Construct scale-invariant feature descriptors for key points in the directional consistency feature map and the aspect ratio continuity feature map. First, extract the local feature region centered on the key point, and the size of the region is proportional to the feature scale of the key point. Then, divide the region into multiple sub-regions, and calculate the feature statistics of each sub-region (such as the gradient direction histogram). Next, perform L2 normalization on the descriptor to improve its robustness to illumination changes. Finally, apply principal component analysis (PCA) to reduce dimensionality, retain the most discriminative feature dimensions, and generate a scale-invariant feature descriptor.

[0141] Verify and optimize the scale invariance of defects. Verify and optimize the scale invariant feature descriptor. First, test the scale invariance of features on artificially synthesized multi-scale defect samples and calculate the feature matching stability index. Then, analyze the variation of features at different scales and establish a scale variation compensation model. Next, adjust the feature extraction parameters through iterative optimization to maximize the scale invariant performance. Finally, apply the optimized feature descriptor to the multi-directional multi-scale coefficient matrix to generate a defect invariant feature map.

[0142] This embodiment solves the problem of scale variation of defect images collected by drones at different distances. The accuracy of defect feature extraction is almost unaffected within the range of 2-10 meters of drone flight altitude variation, and the scale invariance performance is significantly better than the traditional SIFT feature. The feature matching accuracy is improved by about 25%, laying the foundation for the unified expression of features after drones collect images at multiple altitudes.

[0143] According to one aspect of the present application, step S4 is further:

[0144] S41. Multi-scale feature pyramid construction: Based on the defect feature characterization data, a multi-scale feature pyramid network is constructed to capture the defect texture information at different scales and obtain a multi-scale feature pyramid.

[0145] S42, Adaptive defect interference suppression: Analyze the multi-scale feature pyramid, use the adaptive defect interference suppression module to identify and suppress defect noise (such as brick joints, decorative lines, etc.), and generate suppression defect interference features.

[0146] S43, structural texture separation enhancement: Apply the structural-texture separation enhancement algorithm to the interference features of suppression-type defects to improve the distinction between defects and background structures and obtain enhanced separation features.

[0147] S44, Self-attention dense prediction: The enhanced separation features are input into the self-attention dense prediction network to achieve pixel-level defect area segmentation and generate a preliminary defect mask.

[0148] S45, boundary refinement and small defect enhancement: Apply the boundary refinement algorithm based on conditional random fields and small defect enhancement processing to the preliminary defect mask to obtain the final defect candidate area mask.

[0149] like Figure 3 As shown, according to one aspect of the present application, the step of segmenting the defect candidate area under a complex background and obtaining the defect candidate area mask includes:

[0150] Based on the defect feature characterization data, a multi-scale feature pyramid network is constructed to capture the defect texture information at different scales and obtain a multi-scale feature pyramid. Based on the multi-scale feature pyramid, an adaptive defect-like interference suppression module is used to identify and suppress defect-like noise and generate suppressed defect-like interference features. The structure-texture separation enhancement algorithm is applied to the suppressed defect-like interference features to improve the distinction between defects and background structures. Then, the pixel-level defect area segmentation is realized through the self-attention intensive prediction network, and finally the defect candidate area mask is obtained.

[0151] According to one aspect of the present application, based on the multi-scale feature pyramid, the suppression class defect interference feature is further generated as follows:

[0152] Class defect samples and real defect samples are extracted from pre-stored training data, the statistical feature distribution is calculated, a class defect feature statistical model is constructed and combined with a multi-scale feature pyramid, and a conditional random field model is applied to perform context-aware class defect detection to generate class defect detection results; based on the class defect detection results and the class defect feature statistical model, a suppression filter template is constructed to generate an adaptive suppression filter group; the adaptive suppression filter group is applied to the multi-scale feature pyramid to perform structure-preserving feature suppression to obtain the suppression class defect interference feature.

[0153] In one embodiment of the present application, a statistical analysis of class defect features is performed. A multi-scale feature pyramid is received, and a statistical analysis of class defect features is performed. First, typical class defect samples (such as brick joints, decorative lines, pipes, etc.) and real defect samples are extracted from the training data to build a reference feature library. Then, the statistical feature distribution of the samples is calculated, including line width distribution, directional consistency, texture periodicity, and edge gradient characteristics. Next, a multivariate statistical analysis method is applied to identify the most significant feature dimensions that distinguish between class defects and real defects. Finally, a class defect feature statistical model is constructed to describe the feature distribution laws of different types of class defects and generate a class defect statistical feature model.

[0154] Perform context-aware class defect detection. Based on the multi-scale feature pyramid and class defect statistical feature model, context-aware class defect detection is performed. First, the conditional random field model is applied to combine pixel-level features with context information to improve the accuracy of class defect recognition. Then, the matching degree of each candidate linear structure with the class defect model is calculated to generate a class defect probability map. Next, structural regularity analysis is applied to identify line structures with periodicity or geometric regularity (such as brick joints). Finally, local features and global regularity are integrated through multi-level reasoning to obtain class defect detection results.

[0155] Construct an adaptive feature suppression filter. Based on the defect detection results and the defect statistical feature model, an adaptive feature suppression filter is designed. First, specific suppression filter templates are constructed for different types of defect classes, and the filter parameters are dynamically adjusted according to the geometric characteristics of the defect classes. Then, the direction, width, and contrast information of the defect classes are incorporated into the filter design to achieve precise directional suppression. Next, a nonlinear response function is applied to selectively suppress defect-like features while retaining the true defect features. Finally, the adaptive suppression filter group is generated by optimizing the energy function to balance the suppression strength and feature preservation.

[0156] Perform structure-preserving feature suppression. An adaptive suppression filter bank is applied to a multi-scale feature pyramid to perform structure-preserving feature suppression. First, a filter is applied in the feature space to suppress feature responses that match the defect-like pattern. Then, a structure tensor-guided anisotropic diffusion algorithm is used to smooth the defect-like region while maintaining the edge clarity of the real defect. Next, suppression operations are applied at multiple scale levels, and the consistency of the suppression results is ensured by inter-scale consistency constraints. Finally, the suppressed feature pyramid is reconstructed to obtain a preliminary suppressed feature pyramid.

[0157] Feedback correction and optimization enhancement are performed. Feedback correction and optimization enhancement are performed on the preliminary suppression feature pyramid. First, the real defect areas that may be incorrectly suppressed during the suppression process are detected and a correction mask is constructed. Then, a local feature recovery algorithm is applied to selectively restore the defect features that are incorrectly suppressed. Next, contrast enhancement technology is used to improve the distinction between real defects and background. Finally, through multi-scale feature fusion, the suppression results of each level are integrated to maintain the integrity and continuity of the real defects and generate the final suppression defect interference features.

[0158] This embodiment effectively solves the problem of difficulty in distinguishing real defects from defect-like features (such as brick joints, decorative lines, drain pipes, etc.) under complex backgrounds. The false detection rate caused by defect-like interference is reduced, and the accuracy rate under complex texture backgrounds is improved by about 35%. In particular, for masonry buildings and facades with regular decorative textures, the suppression effect is more significant, and the accuracy of distinguishing real defects from defect-like features reaches more than 90%.

[0159] Building facades often have non-defect linear features such as decorative textures, drainage pipes, and window frames. At the same time, the defects themselves may appear intermittent in the image due to surface stains, vegetation occlusion, and other reasons. Existing defect detection algorithms mostly use pixel-level or region-level segmentation methods, which make it difficult to effectively distinguish between real defects and defect-like features, and cannot accurately reconstruct the complete topological structure of intermittent defects, resulting in inaccurate key information such as the connectivity, length, and direction of the defects. There is a lack of a method that can intelligently reconstruct the connections of intermittent defects based on the unique topological characteristics of the defects (such as bifurcation characteristics, stress direction consistency and other physical characteristics) under complex background interference, especially for the limited ability to correctly identify and reconstruct the topology of cross-defect networks. Therefore, it is proposed to Figure 2 According to one aspect of the present application, step S5 is further as follows:

[0160] S51. Defect skeleton extraction: Based on the defect candidate area mask, a topology-preserving thinning algorithm is applied to extract the defect centerline to obtain a defect skeleton map.

[0161] S52. Defect breakpoint detection and classification: Analyze the defect skeleton diagram, identify and classify defect breakpoints, including true end points and false breakpoints, and generate defect breakpoint data.

[0162] S53. Stress field estimation and direction prediction: Combining the 3D model of the building surface and the defect skeleton diagram, the simplified finite element method is used to estimate the surface stress field distribution, predict the defect extension direction, and obtain the stress field direction data.

[0163] S54. Fracture connection under physical constraints: Using defect breakpoint data, stress field direction data and defect feature characterization data, the graphical model reasoning algorithm based on physical constraints generates the optimal defect connection path to form a connection defect skeleton.

[0164] S55. Topology verification and optimization: Apply the topology consistency verification algorithm to the connection defect skeleton, verify and optimize the connection results according to the defect mechanics principle, and obtain a complete defect topology network.

[0165] In one embodiment of the present application, the process of forming a connection defect skeleton is specifically as follows: extracting breakpoint type identification and features. Defect breakpoint data, stress field direction data, and defect feature characterization data are received, and a detailed analysis of the breakpoints is performed. First, the local features of each breakpoint are extracted, including endpoint morphology, directional gradient, and intensity distribution. Then, an unsupervised clustering algorithm is used to divide the breakpoints into multiple types, such as true endpoints, observed occlusion points, and potential connection points. Next, a specific feature descriptor is designed for each type of breakpoint to capture its unique feature pattern. Finally, a type label and reliability score are assigned to each breakpoint to generate breakpoint type feature data.

[0166] Generate candidate paths for breakpoint connections. Based on the breakpoint type feature data and stress field direction data, generate candidate connection paths between breakpoints. First, construct a connection graph network between breakpoints, with candidate connections as edges in the graph. Then, apply directional consistency constraints to filter out connections that do not match the stress field direction. Next, based on the energy minimization principle, generate multiple candidate paths for each pair of possible connected breakpoints, and the path forms include straight line segments, spline curves, and geodesics. Finally, calculate the geometric characteristics and physical rationality score of each candidate path to obtain a set of connection candidate paths.

[0167] Perform physical constraint modeling and path evaluation. Physical constraints are applied to the set of connection candidate paths for evaluation. First, a path evaluation model based on the principles of fracture mechanics is constructed, which includes constraints such as energy release rate, stress intensity factor, and path curvature. Then, the physical score of each candidate path is calculated, and the score reflects the degree of compliance of the path with physical laws. Next, based on additional information such as building structural lines and material boundaries, the path score is adjusted to increase the weight of connections at structurally relevant locations. Finally, a final score is assigned to each candidate path based on geometric similarity and physical rationality to generate path score data.

[0168] Construct a globally optimal connection graph. Combine the path score data and the breakpoint type feature data to construct a globally optimal connection graph. First, model the connection problem as a weighted graph optimization problem, where the nodes of the graph are breakpoints, the edges are candidate connection paths, and the edge weights are based on the path scores. Then, a variant of the minimum spanning tree algorithm is applied to construct the initial connection skeleton. Next, high-scoring connections are iteratively added through a greedy strategy while maintaining topological consistency constraints to avoid crossovers and loops. Finally, a global optimization algorithm is applied to balance local connection quality with overall topological rationality to obtain the initial connection graph.

[0169] Perform graph model reasoning optimization and connection generation. Apply graph model reasoning technology to optimize the initial connection graph. First, construct a Markov random field model to model the connection decision as a probabilistic reasoning problem. Then, define an energy function that includes a single connection quality term and an adjacent connection consistency term. Next, apply the belief propagation algorithm to solve the optimal label configuration and determine the final retained connections. Finally, generate the precise connection path geometry based on the optimization results, merge the connection results with the original defect skeleton, and obtain the connection defect skeleton.

[0170] This embodiment successfully solves the key problem of reconstruction of intermittent defect connections, improves the accuracy of defect connections, and especially improves the correct recognition rate of complex structures such as cross defects and bifurcation defects by about 30%, providing more reliable basic data for defect morphology analysis and cause judgment.

[0171] According to one aspect of the present application, step S53 is further:

[0172] S531, based on the three-dimensional model of the building surface and the building material recognition results, assigning corresponding material mechanical parameters to different areas of the model to generate a building material parameter model;

[0173] S532. Based on the three-dimensional model of the building surface and the parameter model of the building materials, a structured grid suitable for finite element analysis is constructed and load conditions and boundary constraints are defined to generate a simplified finite element model;

[0174] S533, combining the simplified finite element model and the defect skeleton diagram, converting the defect skeleton into a geometric discontinuity in the finite element model, using the extended finite element method to solve the equilibrium equation, and obtaining the surface stress field data;

[0175] S534, calculating the stress intensity factor of the defect tip based on the surface stress field data and the defect skeleton diagram, and generating stress intensity factor data;

[0176] S535. Using stress intensity factor data and surface stress field data, applying maximum circumferential stress criterion and maximum energy release rate criterion, predict the expansion direction of the defect and obtain stress field direction data.

[0177] In one embodiment of the present application, building material property analysis and parameterization are performed. A 3D model of a building surface is received, and material property analysis and parameterization are performed. First, based on the building material recognition results, corresponding material mechanical parameters are assigned to different areas of the model, including elastic modulus, Poisson's ratio, and tensile strength. Then, a nonlinear constitutive relationship model of the material is constructed to describe the mechanical behavior of the material under different stress states. Next, an anisotropic damage model is designed for different building materials such as concrete and masonry to characterize the effect of defects on the mechanical properties of the material. Finally, a building material parameter model containing spatially distributed material parameters is generated.

[0178] Construct a simplified finite element model. A simplified finite element analysis model is constructed based on the 3D model of the building surface and the parameter model of the building materials. First, the 3D mesh model is converted into a structured mesh suitable for finite element analysis, and the mesh is simplified and optimized to maintain a balance between geometric details and computational efficiency. Then, a mapping relationship between node displacement and unit stress is established to construct an overall stiffness matrix. Next, load conditions and boundary constraints are defined to simulate the effects of the building's self-weight, wind load, and temperature changes. Finally, the model parameters are integrated to generate a simplified finite element model.

[0179] Solve the stress field under the influence of defects. Combine the simplified finite element model and the defect skeleton diagram to analyze the stress field under the influence of defects. First, the defect skeleton is converted into a geometric discontinuity in the finite element model, and the extended finite element method (XFEM) is used to deal with the stress singularity introduced by the defect. Then, the equilibrium equations are solved iteratively to calculate the node displacement field and unit stress field. Next, an adaptive mesh refinement algorithm is applied to improve the calculation accuracy in high stress gradient areas such as the defect tip. Finally, the complete surface stress distribution is obtained through post-processing analysis to generate surface stress field data.

[0180] Calculate the stress intensity factor at the defect tip. Based on the surface stress field data and the defect skeleton diagram, calculate the stress intensity factor of the defect tip. First, identify the tip position in the defect skeleton and establish a local polar coordinate system at the defect tip. Then, extract the stress distribution near the tip and apply the displacement extrapolation method or J-integral method to calculate the mode I, II, and III stress intensity factors. Next, based on the defect extension criterion under the mixed mode, evaluate the critical conditions for defect extension. Finally, calculate the energy release rate of each defect tip and generate stress intensity factor data.

[0181] The defect extension direction is predicted based on the energy minimization principle. The defect extension direction is predicted using stress intensity factor data and surface stress field data. First, multiple theoretical methods such as the maximum circumferential stress criterion, the maximum energy release rate criterion, and the minimum strain energy density criterion are applied to calculate the candidate extension directions. Then, a weight model is constructed to integrate the prediction results of multiple criteria in combination with the fracture characteristics of building materials. Next, the Monte Carlo simulation method is used to generate the probability distribution of the extension direction considering the uncertainty of material parameters. Finally, the predicted extension path vector field is generated along the defect skeleton to obtain the stress field direction data.

[0182] This embodiment introduces fracture mechanics theory into the field of defect image analysis, and realizes the scientific prediction of the defect extension direction. It not only provides the physical basis of defect connection, making the connection process conform to the laws of mechanics, but also can predict the future development trend of defects. The accuracy is about 25% higher than that of the method based solely on image features, which provides an important decision-making basis for building safety assessment and preventive maintenance.

[0183] The difficulty of accurately measuring defect width is that existing methods are mostly based on pixel-level calculations, which makes it difficult to achieve sub-pixel precision width measurement. In addition, the optical properties of different building materials have a significant impact on defect edge imaging, resulting in large width measurement errors. Therefore, according to one aspect of the present application, step S6 is further:

[0184] S61. Accurate measurement of defect width: Based on the complete defect topology network and standardized orthophotos, a sub-pixel edge positioning algorithm is used to accurately calculate the defect width distribution and obtain the defect width data.

[0185] S62. Calculation of defect length and distribution: Analyze the geometric characteristics of the complete defect topology network, calculate parameters such as defect length, density and spatial distribution, and generate defect geometric parameters.

[0186] S63. Defect morphological feature extraction: Perform morphological analysis on the complete defect topology network to extract the shape features, bifurcation characteristics and network topology features of the defects to obtain the defect morphological features.

[0187] S64. Defect type and cause analysis: Combined with defect geometric parameters, defect morphological characteristics and building structure information, a multi-feature fusion classification method is used to identify defect types and possible causes, and generate defect type and cause data.

[0188] S65. Defect hazard level assessment: Based on defect width data, defect geometric parameters, defect type and cause data and engineering specifications, assess the hazard level and development trend of defects to obtain defect risk assessment results and defect parameter data.

[0189] According to one aspect of the present application, the steps of extracting and classifying the geometric and physical parameters of the defects to obtain defect parameter data include:

[0190] Based on the complete defect topology network and standardized orthophotos, the defect cross section is extracted: evenly distributed sampling points are set along the defect skeleton, the grayscale or color profile curve perpendicular to the defect direction is extracted, and wavelet transform is applied for multi-resolution analysis to generate defect cross-sectional data;

[0191] Apply high-order derivative filters to the defect cross-sectional data to enhance the defect edge, use a parameterized model to fit the defect cross-sectional profile, achieve sub-pixel edge positioning, and generate sub-pixel edge position data;

[0192] Combining sub-pixel edge position data and building material recognition results, a material-specific edge diffusion model is established to invert the true defect boundary position, eliminate the optical diffusion effect caused by the material, and generate corrected defect width data;

[0193] Based on the corrected defect width data, the defect skeleton is parametrically represented, the discrete width measurement data is mapped to the defect coordinate system, and a robust local regression algorithm is applied to generate a smooth width change curve, identify width mutation points and abnormal areas, and generate a width change curve model that describes the change of defect width along the space.

[0194] Conduct accuracy assessment and anomaly detection on the width variation curve model and corrected defect width data: calculate the statistical uncertainty of the width variation curve model, construct a spatial statistical model of width variation, apply cross-validation technology to evaluate the reliability and stability of width measurement, and generate defect width data;

[0195] The geometric characteristics of the complete defect topology network are analyzed, and the parameters such as defect length, width, density and spatial distribution are calculated. The degree of hazard and development trend of the defects are evaluated based on the defect width data to obtain the defect parameter data.

[0196] In one embodiment of the present application, a multi-resolution defect cross section is extracted. A complete defect topology network and a standardized orthophoto are received to extract the defect cross section. First, uniformly distributed sampling points are set along the defect skeleton, and a sampling line perpendicular to the defect direction is constructed at each sampling point. Then, a grayscale or color profile curve is extracted for each sampling line to capture the local lateral grayscale distribution of the defect. Next, a wavelet transform is applied for multi-resolution analysis to characterize the defect edge features at different scales. Finally, multi-scale information is integrated to generate scale-adaptive defect cross-sectional data.

[0197] Perform sub-pixel edge positioning algorithm. Based on the defect cross-section data, achieve accurate edge positioning at the sub-pixel level. First, apply a high-order derivative filter to enhance the defect edge and improve the accuracy of edge positioning. Then, use parameterized models such as rectangular fitting or Gaussian fitting to fit the defect cross-section profile to achieve sub-pixel edge positioning. Next, calculate the uncertainty of the fitting model parameters and evaluate the reliability of edge positioning. Finally, apply multi-view information fusion technology to integrate observations of the same defect edge in different images, improve positioning accuracy, and generate sub-pixel edge position data.

[0198] Adaptive width correction of building materials is performed. Adaptive width correction of materials is performed by combining sub-pixel edge position data and building material recognition results. First, the influence characteristics of different building materials on defect edge imaging are analyzed, and a material-specific edge diffusion model is established. Then, the real defect boundary position is inverted based on material parameters to eliminate the optical diffusion effect caused by the material. Next, the Bayesian reasoning framework is used to combine prior knowledge and observational evidence to estimate the real width of the defect and its uncertainty. Finally, the corrected defect width data after material correction is generated.

[0199] Fit the spatial variation width curve. Based on the corrected defect width data, a mathematical model describing the variation of defect width along space is constructed. First, the defect skeleton is parameterized and a one-dimensional coordinate system along the defect is established. Then, the discrete width measurement data is mapped to the defect coordinate system to form a width-position data set. Next, a robust local regression (LOESS) or spline fitting algorithm is applied to generate a smooth width variation curve, filtering the measurement noise while retaining the true variation trend. Finally, the width mutation points and abnormal areas are identified, possible special structures of the defect are marked, and the width variation curve model is obtained.

[0200] Conduct width measurement accuracy assessment and anomaly detection. Conduct accuracy assessment and anomaly detection on the width variation curve model and corrected defect width data.

[0201] This embodiment achieves accurate quantification of defect width at the sub-pixel level. The defect width measurement accuracy is improved from the pixel level to the sub-pixel level, and the measurement accuracy is improved by about 3-5 times. The measurement error in the width range of 0.1mm-5mm is controlled within ±0.05mm; at the same time, through material adaptability correction, the influence of different material surface optical properties on measurement accuracy is solved, and the measurement result is closer to the actual physical size of the defect; in addition, through the spatial width change curve, the defect width mutation point and abnormal area can be identified, which provides an important basis for defect cause analysis and hazard assessment.

[0202] According to one aspect of the present application, step S7 is further:

[0203] S71. Multi-level visualization expression: Integrate standardized orthophotos, complete defect topology network and defect parameter data to generate multi-level, interactive visualization expressions and obtain defect visualization models.

[0204] S72. Automatic generation of structured reports: Based on the defect visualization model, defect parameter data and defect risk assessment results, a structured defect detection analysis report is automatically generated, including defect distribution, parameter statistics and risk assessment, to obtain a defect detection result report.

[0205] S73. Historical data comparison and analysis: Compare the current defect parameter data with the historical detection data, analyze the defect development trend, and generate a defect change analysis report.

[0206] S74, Generation of maintenance suggestions: Generate targeted maintenance and reinforcement suggestions based on defect type and cause data, defect risk assessment results and engineering experience database to form a maintenance suggestion plan.

[0207] S75. Inspection quality assessment and feedback: Evaluate the quality of each link in the inspection process, including image acquisition quality, processing accuracy and result reliability, and generate inspection quality assessment data and quality improvement suggestions.

[0208] In summary, the present invention discloses an intelligent detection method for building facade defects, including adaptive acquisition and preprocessing under multiple illumination conditions, geometric correction and registration of multi-view images, extraction and characterization of multi-scale invariant features, segmentation of defect candidate regions under complex backgrounds, reconstruction of defect topological structures based on physical constraints, accurate quantification and classification of defect parameters, and visualization and report generation of results. The present invention solves the technical problems of defect feature enhancement under multiple illumination conditions, complex background interference suppression, unified characterization of multi-view defects, reconstruction of defect topological structures, and accurate parameter quantification through technologies such as regional adaptive illumination compensation, viewpoint invariant feature transformation, class defect interference suppression, physical constraint-based broken connections, and sub-pixel width measurement, thereby improving the accuracy and automation of building facade defect detection.

[0209] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. An intelligent detection method for building facade defects, characterized in that: The following steps are involved: Obtain the original image from the drone, perform preprocessing, geometric correction and registration, and obtain a standardized orthophoto; Perform multi-scale invariant feature extraction and characterization on the standardized orthophoto to obtain defect feature characterization data; Based on the defect feature characterization data, the defect candidate area is segmented under complex background to obtain the defect candidate area mask; Using defect candidate area masks and defect feature characterization data, combined with building mechanics constraints, the defect topology structure is reconstructed to obtain a complete defect topology network; Based on the complete defect topology network, the geometric and physical parameters of the defects are extracted and classified to obtain defect parameter data; Integrate standardized orthophotos, complete defect topology network and defect parameter data to generate defect detection result reports; The steps of reconstructing the defect topology structure in combination with the building mechanics constraints and obtaining a complete defect topology network include: Receive defect candidate area mask, extract defect centerline, and obtain defect skeleton image; Analyze defect skeleton diagram, identify and classify defect breakpoints, and generate defect breakpoint data; Combine the pre-configured building surface 3D model and defect skeleton diagram to estimate the surface stress field distribution, predict the defect extension direction, and obtain stress field direction data; Using defect breakpoint data, stress field direction data and defect feature characterization data, the optimal defect connection path is generated to form a connection defect skeleton; Based on the connection defect skeleton, the connection results are verified and optimized to obtain a complete defect topology network; The steps of segmenting the defect candidate area under complex background and obtaining the defect candidate area mask include: Construct a multi-scale feature pyramid based on defect feature characterization data; Based on the multi-scale feature pyramid, an adaptive defect-like interference suppression module is used to identify and suppress defect-like noise and generate defect-like interference suppression features; The structure-texture separation enhancement algorithm is applied to the interference features of suppression defects to improve the distinction between defects and background structures, and the pixel-level defect area segmentation is achieved through the self-attention dense prediction network, and finally the defect candidate area mask is obtained; The steps of extracting and characterizing multi-scale invariant features and obtaining defect feature characterization data include: Perform multi-directional and multi-scale decomposition on the standardized orthophoto to obtain a multi-directional and multi-scale coefficient matrix; Based on the multi-directional and multi-scale coefficient matrix, the scale-invariant feature descriptor is calculated to generate the defect-invariant feature map; Combining the pre-stored building material recognition results, material-guided feature compensation is performed on the defect-invariant feature map to obtain a compensated feature map; The perspective-invariant feature transformation based on Riemann manifold learning is applied to the compensatory feature map to eliminate the influence of perspective change on the defect feature expression and obtain the perspective-invariant feature representation. It is then combined with the building structure semantic information for context-aware feature enhancement to generate defect feature representation data.

2. The method according to claim 1, characterized in that: The steps of estimating the surface stress field distribution and obtaining the stress field direction data include: Based on the pre-stored building material recognition results, the corresponding material mechanical parameters are assigned to different areas of the building surface 3D model to generate a building material parameter model; Based on the 3D model of the building surface and the parameter model of the building materials, a structured grid suitable for finite element analysis is constructed to generate a simplified finite element model; The defect skeleton diagram is converted into geometric discontinuities in a simplified finite element model, and the equilibrium equations are constructed and solved to obtain surface stress field data; Based on the surface stress field data and defect skeleton diagram, the stress intensity factor at the defect tip is calculated to generate stress intensity factor data; The stress intensity factor data and surface stress field data are used to predict the defect extension direction and obtain the stress field direction data.

3. The method according to claim 1, characterized in that The steps of calculating the scale-invariant feature descriptor and generating the defect-invariant feature map include: Calculate the coefficient differences between adjacent scales for the multi-directional multi-scale coefficient matrix, construct the scale normalization factor and calculate the local extreme points in the scale space to generate a multi-scale differential extreme point set; Based on the multi-scale difference extreme point set and the multi-directional multi-scale coefficient matrix, the defect direction consistency features are extracted and the direction consistency feature map is generated; Combining the directional consistency feature map and the multi-directional multi-scale coefficient matrix, the projection profile analysis is performed and the local curvature and linear measurement of the defect are calculated to generate the slenderness ratio continuity feature map; Taking the key points in the directional consistency feature map and the aspect ratio continuity feature map as the center, extract the local feature area and divide it into a predetermined number of sub-areas, calculate the feature statistics of each sub-area, and generate a scale-invariant feature descriptor; The scale-invariant feature descriptor is verified and optimized, and the defect-invariant feature map is generated by combining the multi-directional and multi-scale coefficient matrix.

4. The method according to claim 1, characterized in that The pre-processing steps include: According to the pre-stored light sensor data and scene reflectivity, the optimal exposure parameters are calculated in real time to obtain the original image of the drone and identify the uneven light areas therein, and generate a light area segmentation mask; Based on the illumination region segmentation mask and the original image of the drone, an adaptive illumination response function model is used to perform differential compensation processing on different regions to obtain an illumination-balanced image. Based on the illumination equalization image, defect-sensitive contrast enhancement is performed to generate preprocessed image data.

5. The method according to claim 4, characterized in that The steps of performing defect-sensitive contrast enhancement and generating preprocessed image data include: Apply a Gaussian derivative filter to the illumination-equalized image, calculate the multi-scale first-order gradient, construct a pixel-level structure tensor matrix for eigenvalue decomposition, and generate a structure tensor feature map; Based on the structural tensor feature map, the main direction feature vector of the defects in the local area is extracted, and a directional Gaussian derivative filter group is constructed to generate a directional enhancement filter group; Apply the directional enhancement filter group to the illumination equalization image to generate a multi-directional response map, apply the non-maximum suppression algorithm to suppress the non-maximum response along the vertical direction of the defect to generate a defect response map, and perform background suppression and defect enhancement processing to generate an enhanced defect response map; Based on the enhanced defect response map, a segmented contrast mapping function is constructed, and a local adaptive histogram equalization algorithm is applied to enhance the local contrast and perform dynamic range compression to generate preprocessed image data.

6. The method according to claim 5, characterized in that The steps for geometric correction and registration include: Combine the pre-stored drone IMU data and the image feature points in the pre-processed image data to calculate the precise camera pose and generate camera pose data; Based on the multi-view images in the preprocessed image data, the geometric model of the building surface is reconstructed to obtain a three-dimensional model of the building surface; Segment and classify the surface area of ​​the three-dimensional model of the building surface and calculate the distance-preserving differential geometry mapping to generate the unfolded surface mapping; Combining the camera pose data, the 3D model of the building surface and the unfolded surface mapping, the pre-processed image data is accurately projected and transformed to obtain an orthographic projection image set; The images of different scales and perspectives in the orthophoto image set are accurately registered and fused to obtain a standardized orthophoto.

7. The method according to claim 1, characterized in that The steps of extracting and classifying the geometric and physical parameters of defects and obtaining defect parameter data include: Based on the complete defect topology network and standardized orthophotos, the defect cross section is extracted and the defect cross section data is generated; Apply high-order derivative filters to the defect cross-sectional data to enhance the defect edge, use a parameterized model to fit the defect cross-sectional profile, generate sub-pixel edge position data, combine it with the pre-stored building material recognition results, establish a material-specific edge diffusion model and invert the real defect boundary position to generate corrected defect width data; Based on the corrected defect width data, a width variation curve model is constructed; Perform accuracy evaluation and anomaly detection on the width variation curve model and the corrected defect width data to form defect width data; The geometric characteristics of the complete defect topology network are analyzed, the defect length, density and spatial distribution parameters are calculated, and the defect parameter data are obtained.

Images